v0.9.0 Release Notes

Released 2026-03-06

v0.8.0 -> v0.9.0

Summary

Adds file injection, image input, error memory, and web search caching as new user-facing capabilities, alongside a remove command and public API refactor that exposes session/agent selection for library consumers.

翻譯

新增檔案注入、圖片輸入、錯誤記憶與網路搜尋快取等使用者功能，同時新增 remove 指令並重構公開 API，將 Session / Agent 選取邏輯開放給函式庫使用者直接呼叫。

Changes

FEAT

• Add --file <path> flag to inject local file content into prompts as embedded text blocks
• Add --image <path> flag with multi-part message encoding for vision-capable providers
• Add remember_error and search_errors tools for persistent cross-session error memory
• Add file-based web search cache with SHA-256 keying and 1-hour TTL to reduce redundant DDG requests
• Add remove command for interactively removing configured providers from the keychain
• Split list into list (models) and list skills subcommands for clearer output

翻譯

• 新增 --file <path> 旗標，將本機檔案內容以嵌入文字區塊形式注入 prompt
• 新增 --image <path> 旗標，支援視覺模型的多部分訊息編碼
• 新增 remember_error 與 search_errors 工具，實現跨 Session 的持久化錯誤記憶
• 新增以 SHA-256 為 key、TTL 一小時的檔案型網路搜尋快取，減少重複 DDG 請求
• 新增 remove 指令，互動式移除已設定的 Provider
• 將 list 拆分為 list（模型）與 list skills（技能）子指令，輸出更清晰

REFACTOR

• Expose GetSession, SelectAgent, SelectSkill as public API; add Name() method to Agent interface
• Set HTTP timeouts on browser fetch and API requests to prevent indefinite hangs
• Truncate oversized tool results before appending to session messages to avoid context overflow
• Rename ExecData.Input → ExecData.Content, ExecData.Images → ExecData.ImageInputs; add ExecData.FileInputs
• Deduplicate repeated identical tool calls via alreadyCall cache map in the execution loop

翻譯

• 將 GetSession、SelectAgent、SelectSkill 開放為公開 API；在 Agent 介面新增 Name() 方法
• 為瀏覽器擷取與 API 請求設定 HTTP 逾時，防止無限等待
• 在將工具結果附加至 Session 訊息前截斷過大內容，避免 context 溢出
• 將 ExecData.Input 重新命名為 ExecData.Content、ExecData.Images 為 ExecData.ImageInputs，並新增 ExecData.FileInputs
• 在執行迴圈中以 alreadyCall map 去重重複的工具呼叫

UPDATE

• Refine agent/skill selector prompts to produce more deterministic routing decisions
• Simplify system prompt rules to reduce token usage while preserving behavioral constraints
• Support multi-word CLI input via joined args with \n escape support; add run-allow subcommand
• Increase conversation history limit and improve summary prompt formatting

翻譯

• 優化 Agent / Skill Selector 提示，提升路由決策的確定性
• 精簡系統提示規則，降低 token 用量，同時維持行為約束
• CLI 輸入支援多詞 arg 自動串接與 \n 跳脫；新增 run-allow 子指令
• 提高對話歷史保留上限，改善摘要提示格式

---

Files Changed

---

Generated by SKILL
Full Changelog: https://github.com/pardnchiu/agenvoy/compare/v0.8.0...v0.9.0